Yeah I tried to have a conversation about security yesterday with Claude and it immediately stopped me - I was taken aback, I didn’t expect it at all.

This is highly problematic.

> (I did apply for OpenAi's Cyber Pilot but got no response)

Same experience with Anthropic's. I applied for my employer, and.. 0 response.

I’m in the exact same boat, although Codex isn’t quite as bad.

Fable smacked me for asking it to design a secure app without obvious security flaws and to double check it wasn’t using libraries with known security problems.

How much do both cost together? Claude Code alone is 200 bucks a month... I'm satisfied with the 100 buck subscription for now with headroom

Was it applying to the program for an organization or individual? The description for the individual application page makes it sound pretty straightforward compared to getting access for an organization.

The product strategy of 'consumer-grade' AI making deliberately insecure software, and then selling you limited access to the model that can fix it (if they think you deserve to pay them) is just diabolical.

(Grade 3 AI which can hack both previous tiers is exclusively sold to the highest bidder.)

> K3 is the only frontier model I can have a serious conversation with about my product's security.

This is wrong IMO. You should have a serious conversation about your products security with someone who is actually trained on that subject. LLMs are useless if you don't already know more about the thing than the LLM, or if you don't care too much about the outcome (internal tools etc.)

> I have no idea why HN still treats them like their the ethical good guy.

I don't know that HN usually does. But they did buy a decent amount of publicity by (initially) telling the US government no in using the models for war.

There are plenty of people (myself included) who are critical of Anthropic. Sentiment isn't universal amongst a crowd, it might just feel that way due to the way comments are surfaced on HN once posts reach the front page.

They spend $40mil for lobbying, I'm sure they can also spare some millions to this place (and other places like reddit). They all do.

And what else, forbid running workloads outside of US so you can’t offshore workloads? What’s the rationale behind policies that can be detrimental only to US companies?

> I have no idea why HN still treats them like ~their~ they are the ethical good guy.

s/good/least bad/

Which company would you put higher?

*squishy

These AI CEOs are psychopaths, they will do anything to get big bucks from defense contracts. There is recent interview with Dario where he said Iran school bombings do not even violate their redlines[1]. Full interview [2] shows that he think like Trump (we are the superior nation)

1. Breaking Points Discussion: https://www.youtube.com/watch?v=SIUrshGwgDI

2. Full Bloomberg interview: https://www.youtube.com/watch?v=x2VHFgyawPE

The future will probably have most of all companies running local models, simply because the alternative would be essentially every company that uses LLMs ending up becoming completely dependent upon Anthropic et al. And that dependence would be milked to the point of absurdity once solidly established.

And as a more general point - more major competitors in a domain is very good for everybody except those competitors themselves, who would rather there be as few as possible.

Nvidia is subject to export restrictions on GPUs to China. So Chinese labs have responded by allowing American hosts to run their open-weight models in the US, where there are no restrictions. We see this in the EU too where Scaleway, and now Hetzner are getting into the inference host space serving Chinese open-weight models.

So Nvidia benefits from more inference providers running many Chinese models that they otherwise would not have been able to service.

NVIDIA sells the tools. The more the better it is for them.

Microsoft and Meta are also-rans at this point. Their best hope of catching up is probably leveraging their infra and open models.

Different companies have different motivations. I'm sure some are genuine, but a lot of them simply lost the race (to OpenAI/Anthropic/Google) and know they can't compete anymore, so are shifting strategies. How many times has Microsoft done exactly this in the past?

They've seen the government officials parroting OpenAI and Anthropic lobbies talking points, and they have smart enough people to see the signs of upcoming regulatory capture.

If I were cynical, I'd say it isn't helped by the current US administration.

They are worried about OpenAI and Anthropic making their own processors and controlling the whole stack is my guess.

If there are effectively only two AI companies allowed to train or run model inference then demand for hosting, GPUs and data centers goes down, so they'd be able to squeeze their suppliers that much more too.

Probably a real effort to push for some kind of sanctions or commerce block on Chinese models. Anthropic just doubled their political spending to $40 mil for the midterms to push for "AI Safety."

Almost certainly these companies are using similar strategies as the Chinese models which they fear will be made illegal.

It's also the case that it makes it hard to attract customers if your openweight model is banned. A major reason the likes of Qwen, Kimi, GLM, and Deepseek are popular (well, at least highly talked about) is because of the open weight models they gave away.

Nvidia benefits no matter what because their hardware is being used. Microsoft and Meta benefit by making sure that the gap between "Anthropic/OpenAI" and "everyone else" doesn't widen, or they'll be hopelessly dependent on those frontier labs.

Probably the fact that they don't want Anthropic or Google or OpenAI to have access to their data theoretically, and that they do want to use good AI, and that they don't want to spend the money themselves to make a good model...

I think Meta would love nothing more than to throw a very large wrench into the works of OpenAI/Anthropic. Open models could mean one or both go bankrupt which is good for them.

Nvidia sells the hardware.

Microsoft sells the hosting.

They're trying to bring down Anthropic. There's been a huge smear campaign happening all year.

Yeah, this letter is just proof of (the by now long standing fact) of corporate power/corporate capture of American governance. Tons of popular protests against big AI? Makes no difference. Joint corporate letter about big AI? Possibly the government will do something.

> Equally telling is the list of companies that did not sign it -- including Google, Amazon.

Could you elaborate? I don't understand your implication.

Amazon in particular hosts open weight models via Amazon Bedrock (e.g. Ministral 14B 3.0, DeepSeek V3.2, ...)[1].

In case it changes the context of this comment, I am a Meta employee, all opinions are my own.

[1] https://docs.aws.amazon.com/bedrock/latest/userguide/model-c...

Google and Amazon are Anthropic investors right?

DeepSeek has published some really good papers. Lately they're pushing really hard for dramatically cheaper serving costs.

I may be remembering wrong, but reasoning was first demonstrated by DeepSeek. Edit: I am indeed remembering wrong, seems o1 was first.

To some degree even unintentionally this is all inevitable. I know this isn’t what you’re taking about, but it’s a similar line of thought. These models are consuming public free information, and they’re all also producing free public information, so they’re all pissing and drinking into same pool.

If we go down the line of dead internet theory which I’m becoming more convinced of these days, the volume of information that’s not necessarily original or extracted from reality and just interpolated and extrapolated from existing information in different ways by LLMs will greatly outnumber human information coming up.

In which case these models should.. start to converge on the same data I imagine, with slightly different behaviors within that. One big generative orgy feedback loop.

There are a ton of DeepGEMM kernels in all of the flashinfer adjacent stacks including trtllm. Bunch of the headers still have the CLion pragmas and shit.

Many of the Chinese optimizations are public because they are published by the Chinese labs themselves. Hard to say what the labs are doing of course.

Let me put it this way: if they're not ripping off all that they can, their investors need to shitcan the leadership.

Makes sense Microsoft is placing it self as a model host and os providing open weight models for much cheaper on their infra already.

No wonder openai and anthropic are terrified

[dead]

Ideology? I think this administration is the least captured by ideology in the last few decades. It's mostly brazen self-interest, for both entertainment & profit.

> Start with the outcome you believe will be the most in line with the spirit and traditions of the open source community

there's nothing open source about an open weight model. It's more like the binary blobs Linux fought so hard against. If you can't compile, or in this case train, from source then you're missing the "source" part of open source.

Too bad Microsoft couldn’t see that before they spent the money…

> I'm not sure how else to say it other than that it's honestly just embarrassing to watch China and others run circles around us in the USA.

The gap is shrinking, but considering their models still lag our frontier, I wouldn't say they "run circles" around us.

i'm glad someone else sees this too. There's nothing open source about an open-weight model. I feel like they're just trying to associate themselves with open source and the positive impacts it has had on software technolgoy.

It will be like DeCSS. Now imagine models (links) in QR code spread around, or steganographied in BluRay MKVs. So much creative disobedience.

…and with the size of each class determined by common target memory sizes i.e. 32GB or 64GB or 128 GB minus cache and a little overhead

They already do and everyone has been using that concept for years now, it's known as RLAIF.

I don’t think the Sinoloa cartel cares at all about frontier models, and the military already runs their own locally deployed editions of Mythos.

> that day is over.

I see zero indications of this. As a matter of fact, I see evidence of the opposite, the Chinese government has just come out in support of open weight AI last week.

What does copyright have to do with anything? You also cannot copyright an encryption algorithm, but the US Government banned essentially mathematics for YEARS[0].

[0] https://en.wikipedia.org/wiki/Export_of_cryptography_from_th...

You can randomly shift every pixel in a movie by +/-1, doesn't make the original not copyrightable.

I think if anyone at Apple signs a plea that has the word 'open' in it they spontaneously combust into flames

I wonder if it has anything to do with Apple’s legal case against OAI.

This would support the hypothesis that they are incentivized as compute providers and not simply because they're losing the frontier llm race

If it doesn't involve a walled garden subject to Apple's domination and control, it's not an Apple product.

Apple voted long time ago not to participate their contribution was giving OpenAI nothing and giving Google a $1 billion refund on the default search position money.

Maybe less shocking, but still, so is google. I know they are one of the labs, but that being said they rely far less on gemini for revenue as opposed to the big 2.

Apple is allergic to industry-spanning consortiums.

Apple is not an enterprise tech company. It makes and sells luxury consumer devices.

Apple would be rather unaffected since they already signed a deal with Google/Gemini.

Because their long-term goal is to make AI a walled garden subscription.

Both the CEO of Palantir and the CEO of Microsoft have recently criticized very strongly the business model of OpenAI and Anthropic.

Therefore it is no surprise that both companies are among the signatories.

The motivation of Palantir is very transparent, they have launched a product created in cooperation with NVIDIA, which is a turnkey system (Sovereign AI) that includes all hardware (i.e. a rack with servers full of NVIDIA GPUs) and all software needed by a company to run on its premises LLM inference and also training/fine tuning.

Thus the Palantir product competes with OpenAI and Anthropic and it benefits from the existence of open weights LLMs.

The motivation of Microsoft is more obscure for now, but because its criticism was very similar to that of Palantir, about how dangerous it is for corporations to delegate the processing of their AI needs to external entities like OpenAI and Anthropic, I assume that Microsoft is also going to propose an alternative to using remotely the OpenAI and Anthropic APIs.

Unlike Palantir, Microsoft is not likely to suggest self-hosting on premises, but I suppose that they might promote self-hosting on some kind of instances rented from Azure.

Commoditization of the model layer is to Azure's benefit. It would help to avoid the model layer becoming a platform that commoditizes cloud into a pure bare metal business. It also redistributes 'consumer' surplus back to the end user via a lower effective price, which would conceivably increase demand for IaaS and/or PaaS which are complements.

My hunch is they are all seeing the problems at OpenAI and some other collateral damage on the horizon. They don’t want the US government intervening to try to save OpenAI. A little controlled burn is probably good for the industry long term.

> But the letter is really about _American_ open weights models.

AI models don’t really have nationality. They don’t have race/ethnicity fields on their model cards. It’s impossible to determine the country of origin of a model by looking at their weights. There is no DNA test for AIs. They exist in a mathematical space where human-made borders make no sense.

This means IF there is a regulation for open weight models, it has to apply to ALL open weight models. There is no any other option, since you can always train/fine-tune a model to change it’s weights and rebrand it as a new model.

It helps to understand there are multiple factions within the white house and trump admin. One faction is very hawkish towards allies and enemies, especially China. Another is the one sending ICE to invade American cities, the Christian nationalism and such. Another is very focused on wealth expansion for friends and family. There is an amount of Venn to these groups.

Signatories appear to group into

(a) sellers of hardware (Nvidia, IBM) and those who rent hardware out to others (Amazon, Microsoft)

They want open models because they don't have to license it to run it on their hardware, so they can offer lower cost and inspect what they are offering to clients

(b) open model creators (Arcee, Perplexity, Mistral, IBM, Microsoft)

...who would be unable to work if open-weight models are banned

(c) Software-heavy companies ( Mozilla, The Linux Foundation , ServiceNow, Microsoft, IBM, heck all of them)

...who want models that are cheap to run because that reduces the cost of running an LLM (whether you are trying to replace a software developer or just assist them, the value of a reduced cost LLM is directionally the same)

> Microsoft is somewhat unexpected to me

MS is playing literally all the sides in this whole AI thing. They are a massive investor in oAI, and have lots of rights to use their models for a long time. They also have the cloud where they offer inference w/ all the benefits of already being there, data retention, etc. They also have a research wing that can build models (rumours are they're aiming for frontier-ish). I don't think there's currently an angle in this AI thing that MS hasn't aimed for.

> The call to action is all about keeping the American open weights ecosystem competitive. I didn't even realize that was under threat, so I feel like I'm missing something.

OpenAI and Anthropic can bring American open-weight model providers like Meta to court but not the Chinese. Chinese companies breaking American IP law (whether the law is correct or not) are basically immune. This has been going on forever, I guess enough is at stake finally for the federal government to care.

Google and Amazon both have incredibly large compute deals with frontier labs and GovCloud. They may not want to sour on that cash.

The important thing is that Sam Altman and Dario Amodei become trillionaires. Everything must be decided in service of that one primary goal.

And the verdict is still out on if it's going to be cheaper than employees in the long run. Maybe for companies in HCoL areas (SV/NYC), but any smaller mid-market org in a LCoL area, frontier model token spend may not end up all that cheaper when your employee salaries are $55k-$75k/year so you're mostly evaluating it as an additional force multiplier expense rather than a headcount replacer.

What is a "Chinese model"?

If I change one weight in Kimi, is it still a Chinese model?

If I fine-tune Kimi on pro-America nationalistic freedom loving anti-Chinese propaganda, is it still a Chinese model?

If I distill Kimi from one server to the next without ever directly transferring any weights, is it still a Chinese model?

If Claude accidentally trains on some of Kimi's output, is Claude then considered to be a Chinese model?

If China's next open-weight model is released secretly through a European company, is it still a Chinese model?

> ...the issue is Chinese models, whether open or not, and that's what's looking to be banned...

Is the US planning to invade China and take away their models? Because otherwise China is still going to have the models even if the US bans them. I don't think they're going to be going with the US position on this one.

The conversation going on in the industry is a bit broader than that. OpenAI's head of strategic futures just said last week that open models are inherently decelerationist, ungovernable, and will slow development on the frontier. You'll notice OpenAI does not appear as a signatory here despite their relationship with Microsoft.

Why anyone would care what Hegseth thinks about open weight models? What's wrong with Chinese models? The censorship? I think it's good to have diversity of opinion. Get US flavored output on sensitive subjects from US models and ZH on ZH models. That's a useful tool.

They did. Check out the Phi models (e.g., https://huggingface.co/microsoft/phi-4).

Ok, but please don't fulminate or post unsubstantive comments here.

https://news.ycombinator.com/newsguidelines.html

It's not worthless, you don't have to agree with P on everything, or make it personal - it works as strictly business - only where and when appropriate.

Regardless of the motivations behind the paper, the fact is that they are right. It's a topsy-turvy world where America closes up and China opens, and I hope it gets corrected soon.

And besides, it seems to me like the ethical choice. Given how these models are, in some very real sense, mechanical plagiators, built on the generosity of creators past and present, some of them now in danger of being replaced by the machine. I think the least the labs can do is open these models up. These and other such considerations were the reason OpenAI started with that name. Of course, it was questionable that those ideals would survive the encounter with generational wealth. Just look up what the founders of Google were saying about advertising when they were two students tinkering at an as of yet unproven tech. Same thing for OpenAI, self-interest speaks that much louder when there's real money on the table.

It just boggles the mind that people now make excuses for their all-too-predictable about-turn.

This is some bizarre victim inversion. The providers of closed models are the ones who are trying to use regulation to stop their open model competition, not the other way around.

I'm very confused by this comment, I don't know who you're referring to.

ha! who is groveling to the white house to stop competition? openai, anthropic and google, they are the losers. if they want to compete, compete.

I don't think that is what is happening.

Instead a bunch of tech companies are gathering to try to stop OpenAI and Anthropic fear-bouncing the White House and the Republican Congress into giving them regulatory capture and repeating the mistakes they are making around RISC-V.

Those mistakes won't just entrench two companies, they will entrench the bigger-better-faster-more model (closed companies making ever bigger cloud-bound models) when it is abundantly clear that enormous progress can still be made on smaller, even desktop-bound models (where, due to distribution, open weights are essentially inevitable).

Regulatory capture that stops open weights work will also have impacts on local and on-device AI work, as well as on academic research.

You can call out regulatory capture evwn when you would do the same if the shoe were on the other foot.

I thought the whole point of open software development was anyone could take your thing and improve it.

It would seem as if the community either isn't doing that or is relying on the Chinese to do that.

I’ll never forgive Microsoft for their hostile actions against open source and open standards in the 90s. I’ll never believe any endorsement from them of open-anything.

Yeah, the endorsement logos at the bottom are majority bad actors with regards to openness in code. Many are bad actors with regards to openness in _society_.

Yeah, never ask a fox to count your chickens.